Skip to content

build: bump TransformerEngine to release_v2.14 (dev) - #4332

Merged
ko3n1g merged 11 commits into
NVIDIA:devfrom
ko3n1g:ko3n1g/bump-te-release-v2.14-dev
Apr 17, 2026
Merged

ko3n1g merged 11 commits into
NVIDIA:devfrom
ko3n1g:ko3n1g/bump-te-release-v2.14-dev

Conversation

@ko3n1g

@ko3n1g ko3n1g commented Apr 16, 2026 •

Copy link
Copy Markdown
Contributor

Summary

Bumps the pinned TransformerEngine commit from 5671fd36 to f031cf87 — the HEAD of the release_v2.14 branch.

  • pyproject.toml: updated [tool.uv.sources] rev
  • uv.lock: regenerated inside the megatron-lm:local container

Also adds retry logic (3 attempts, 10 s backoff) to all apt-get update && apt-get install steps in CI to handle transient Ubuntu mirror hash-sum mismatches that were causing intermittent failures.

Marks test_fine_grained_activation_offloading_with_cuda_graph as flaky_in_dev while the underlying issue is investigated (tracked in #4350).

Disables 5 MoE functional tests that fail on num-zeros metric comparison after the TE v2.14 bump:

Changes

File Change
pyproject.toml TE bumped to release_v2.14 head (f031cf87)
.github/actions/action.yml Install uuidgen step wrapped in retry loop
.github/workflows/cicd-main.yml Both Install GH CLI steps wrapped in retry loops
tests/unit_tests/pipeline_parallel/test_fine_grained_activation_offloading.py test_fine_grained_activation_offloading_with_cuda_graph marked flaky_in_dev
tests/test_utils/recipes/h100/moe.yaml 5 test cases disabled (-broken scope)
tests/test_utils/recipes/gb200/moe.yaml 5 test cases disabled (-broken scope)

Example diff

# pyproject.toml
-transformer-engine = { git = "...", rev = "5671fd3675906cda1ade26c24a65d3dedd88eb89" }
+transformer-engine = { git = "...", rev = "f031cf87bd054c7558b887df7bed93975456667f" }
# apt-get retry pattern applied to all install steps
- name: Install uuidgen
  shell: bash -x -e -u -o pipefail {0}
  run: |
    for i in 1 2 3; do
      apt-get update && apt-get install -y uuid-runtime && break
      echo "apt attempt $i failed, retrying..."
      sleep 10
    done

Signed-off-by: oliver könig <okoenig@nvidia.com>
@copy-pr-bot

copy-pr-bot Bot commented Apr 16, 2026

Copy link
Copy Markdown

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

@ko3n1g

ko3n1g commented Apr 16, 2026

Copy link
Copy Markdown
Contributor Author

/ok to test c2a8f74

@svcnvidia-nemo-ci svcnvidia-nemo-ci added this to the Core 0.16 milestone Apr 16, 2026
@ko3n1g
ko3n1g marked this pull request as ready for review April 16, 2026 09:23
@ko3n1g
ko3n1g requested a review from a team as a code owner April 16, 2026 09:23
ko3n1g and others added 2 commits April 16, 2026 10:16
Transient mirror inconsistencies can cause apt-get to fail with a Hash
Sum mismatch. Wrap the update+install in a 3-attempt retry loop so the
step self-heals without manual re-runs.

Signed-off-by: oliver könig <okoenig@nvidia.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: oliver könig <okoenig@nvidia.com>
Same transient mirror hash-mismatch fix as applied to action.yml.
Both "Install GH CLI" steps now retry up to 3 times before failing.

Signed-off-by: oliver könig <okoenig@nvidia.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: oliver könig <okoenig@nvidia.com>
@ko3n1g

ko3n1g commented Apr 16, 2026

Copy link
Copy Markdown
Contributor Author

/ok to test

@ko3n1g

ko3n1g commented Apr 16, 2026

Copy link
Copy Markdown
Contributor Author

/ok to test

- test_paged_stashing.py::TestPagedStashing::test_forward_backward_4_layers
  (tracked: NVIDIA#4339)
- test_vision_cuda_graphs.py::TestVisionTECudaGraphHelper::test_create_cudagraphs_no_callables_is_noop
  (tracked: NVIDIA#4340)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: oliver könig <okoenig@nvidia.com>
@ko3n1g
ko3n1g requested review from a team as code owners April 16, 2026 14:02
@ko3n1g

ko3n1g commented Apr 16, 2026

Copy link
Copy Markdown
Contributor Author

/ok to test

…r_budget as flaky_in_dev

Fails with TE v2.14: TEGroupedMLP fused impl not supported for
moe_paged_stash=True + use_transformer_engine_op_fuser=True.
Tracked: NVIDIA#4345 (same root cause as NVIDIA#4339)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: oliver könig <okoenig@nvidia.com>
@ko3n1g

ko3n1g commented Apr 16, 2026

Copy link
Copy Markdown
Contributor Author

/ok to test

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: oliver könig <okoenig@nvidia.com>
@ko3n1g

ko3n1g commented Apr 16, 2026

Copy link
Copy Markdown
Contributor Author

/ok to test

Signed-off-by: oliver könig <okoenig@nvidia.com>
@ko3n1g

ko3n1g commented Apr 16, 2026

Copy link
Copy Markdown
Contributor Author

/ok to test bf1bd52

…s flaky_in_dev

Tracks: NVIDIA#4350

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: oliver könig <okoenig@nvidia.com>
…ph as flaky_in_dev

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: oliver könig <okoenig@nvidia.com>
@ko3n1g

ko3n1g commented Apr 16, 2026

Copy link
Copy Markdown
Contributor Author

/ok to test

Tests fail on num-zeros metric comparison (exact and approximate) after
bumping TransformerEngine to release_v2.14. Disabling until golden values
are updated or the root cause is fixed.

Affected test cases:
- gpt3_mcore_te_tp2_pp1_te_8experts2parallel_overlap_grad_reduce_param_gather_groupedGEMM
- gpt3_mcore_te_tp2_pp1_te_a2a_ovlp_8experts_etp1_ep4
- gpt3_mcore_te_tp2_pp1_te_8experts2parallel_ddp_average_in_collective
- gpt3_moe_mcore_te_tp4_ep2_etp2_pp2_resume_torch_dist_dist_optimizer

Tracked in NVIDIA#4352.

Signed-off-by: oliver könig <okoenig@nvidia.com>
@ko3n1g

ko3n1g commented Apr 16, 2026

Copy link
Copy Markdown
Contributor Author

/ok to test

… bump

Signed-off-by: oliver könig <okoenig@nvidia.com>
@ko3n1g
ko3n1g merged commit ac6ca5b into NVIDIA:dev Apr 17, 2026
226 of 229 checks passed

This branch was previously deployed

1 inactive deployment
test — 75642b70 Deployed Apr 17, 2026 by copy-pr-bot[bot] via cicd-wait-in-queue #11668
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants